test(rccl): Validate GIN proxy poll batching and the RMA aggregation hint by rahulvaidya20 · Pull Request #13221 · ROCm/rocm-systems

rahulvaidya20 · 2026-10-09T22:46:44Z

Motivation

The GIN host proxy drains up to NCCL_GIN_PROXY_POLL_BATCH GFDs per peer on each progress tick and passes ncclRmaOptFlagsAggregateRequests to the RMA backend when that peer's next operation is already queued. Neither was documented, the RMA example plugin could not serve as a GIN proxy backend, and nothing showed the hint reaching a backend end to end.

Technical Details

Extend the GIN host proxy microtests from test(rccl): add host microtests for the GIN host proxy #13596 (rccl-UnitTestsMicroGinHostProxy) with two cases: the hint is set on put-with-signal, get and VA-signal ops, not only plain puts, and puts queued ahead of an iflush are hinted while the flush ends the run.
RMA example plugin: report NCCL_NET_DEVICE_GIN_PROXY so NCCL_GIN_TYPE=2 keeps the proxy over it, count data operations and hinted ones and log them at finalize, and pin its log levels to RCCL's numbering at compile time.
Add GinMPIDeviceTests.PutBurst_LocalCounter: 32 threads issue 256 puts to one peer and wait only on a local counter, so it also runs over a backend that moves no data. It skips on non-proxy GIN backends.
Add ext-rma tests: one runs that burst through the RMA example at poll batches 1, 4 and the default and checks the hinted counts stay within each batch's bound; the other checks, without GPUs, that iput, iputSignal and iget each count a hinted op.
Run the burst at the default poll batch and at a batch of one in the gin_device_get_flush groups of the MI300X and MI355X functional test-runner configs.
Document NCCL_GIN_PROXY_POLL_BATCH and when the GIN proxy sets the aggregation hint.

Issue Tracking

JIRA ID: AICOMRCCL-2018

Test Plan

Run the full host-test pipeline (run_host_tests.sh configure, build, run).
Run PutBurst_LocalCounter and WaitCounter_Local over the built-in IB GIN proxy backend at poll batches 1 and 32, and check they skip under NCCL_GIN_TYPE=7.
Run the full ext-rma suite, plus 20 repeated default-batch bursts over the RMA example plugin.
Build the RMA example standalone and in-tree with -DBUILD_PLUGIN_EXAMPLES=ON.

Test Result

Host pipeline: passed with no failures, including the 75 GIN host proxy microtests.
PutBurst_LocalCounter and WaitCounter_Local passed over the IB backend at both batches on MI300X (gfx942, InfiniBand, ROCm 10.0) and MI350X (gfx950, RoCE, ROCm 10.1), and skip under NCCL_GIN_TYPE=7.
ext-rma: 7/7. The RMA example received all 256 puts, with 0, 192 and 248 hinted at poll batches 1, 4 and 32, and 248 in 20 of 20 repeated default-batch runs.
Both RMA example builds succeeded.

Submission Checklist

I have read and followed the contributing guidelines

therock-pr-bot · 2026-10-09T22:47:40Z

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

Copilot

🟢 Approval recommended

The tests, example instrumentation, build wiring, and documentation consistently validate the intended behavior.

0 open findings

What changed in this PR

Adds end-to-end validation and documentation for GIN proxy poll batching and RMA aggregation hints.

Changes:

Adds host microtests and an MPI burst test covering batching and hint propagation.
Enhances the RMA example plugin with GIN proxy support and operation counters.
Adds integration coverage and documents the related environment variable and hint semantics.

File	Description
`projects/rccl/tools/scripts/test_runner/configs/ci-precheckin.json`	Updates microtest description.
`projects/rccl/test/transport/GinDeviceMPITests.cpp`	Adds the local-counter put burst test.
`projects/rccl/test/test_categories_micro_gin_host.yaml`	Expands microtest category documentation.
`projects/rccl/test/host/gin-host-proxy-test.cc`	Tests batching and aggregation hints.
`projects/rccl/test/host/CMakeLists.txt`	Builds the new host microtests.
`projects/rccl/test/ext-plugins/tests/ext-rma/test_rma_plugin.py`	Adds end-to-end RMA batching tests.
`projects/rccl/plugins/rma/example/plugin.c`	Adds GIN proxy reporting and counters.
`projects/rccl/plugins/rma/example/nccl/net_device.h`	Defines network-device types for the example.
`projects/rccl/plugins/rma/example/nccl/common.h`	Aligns logging levels with RCCL.
`projects/rccl/docs/how-to/using-rccl-rma-plugin-api.rst`	Documents aggregation-hint semantics.
`projects/rccl/docs/api-reference/env-variables.rst`	Documents the poll-batch setting.

🧠 Review effort: Balanced

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

amd-argus

ARGUS

Findings

1 finding(s) that could not be posted as inline comments

projects/rccl/tools/scripts/test_runner/configs/ci-precheckin.json:257 (Low) Two descriptions of rccl-UnitTestsMicroGinHost still name only gin_host.cc after the binary gained a second unit The per-test description at configs/ci-precheckin.json:257, the one the runner prints, still reads "Host-only gin/gin_host.cc microtests" and names only GIN_PROXY_NTHREADS, while :318 was updated. test/host/MICROTEST_README.md:251 is stale the same way: it is the standing record of units under test and still lists this binary as covering gin/gin_host.cc alone, where every other entry there enumerates all its units. Could both pick up gin_host_proxy.cc and GIN_PROXY_POLL_BATCH?

Bottom line: Worth settling before approval: _unit_tests_mpi probes only <RCCL_INSTALL_DIR>/build/..., which the convention the repo's own CI exports never matches, so the new ext-rma test always skips.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

rahulvaidya20 · 2026-10-10T00:48:48Z

Addressed the Argus review in 4 commits:

_unit_tests_mpi lookup (2a8caa3fbd): also looks under the source tree's build/{release,debug}/test, so the burst test finds the binary when RCCL_INSTALL_DIR points at <prefix>/lib. The skip message now says --debug --enable-mpi-tests.
Unbounded stream sync (4fee7ca7cf): PutBurst_LocalCounter uses syncStreamWithinTimeout(stream, 60), so a proxy stall fails instead of hanging.
iflush coverage (bd55439c28): new FlushEndsAHintedRun host microtest. It checks that the puts before an iflush are hinted and that a flush with no request completes without being tested. The same commit updates the stale rccl-UnitTestsMicroGinHost descriptions in ci-precheckin.json and MICROTEST_README.md.
Nits (1a3a8a5762, bd55439c28): NCCL_NET_DEVICE_INVALID_VERSION, log call indentation, and loop braces.

Kept the lower bound of one hinted op in the ext-rma test. It is the only end-to-end check that the hint reaches the plugin; the host microtests already pin the exact behavior deterministically. A comment now notes it assumes the burst outpaces the proxy, which a full CTA against the 32-slot queue does: 20 of 20 default-batch runs gave 248 hinted ops.

Re-validated on MI350X (gfx950, RoCE, ROCm 10.1): host microtests 26/26, ext-rma 6/6 (hinted ops 0, 192 and 248 for batches 1, 4 and 32), the burst test with RCCL_INSTALL_DIR set to the library directory, and PutBurst_LocalCounter plus WaitCounter_Local over the built-in IB verbs GIN proxy.

amd-argus

ARGUS

Bottom line: Looks approval ready. Everything from the last round is addressed and nothing above Low survived verification.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

rahulvaidya20 · 2026-10-11T07:27:35Z

Second Argus round:

Signal region in the host fixture (2fa68ba3bb): nSignalsPerContext is now 2, so the signal ID the fixture queues falls inside the region. Host microtests 26/26.

Not changed:

Below-1 clamp untested: the clamp lives in ncclGinProxyCreateContext, which needs a real collComm, RMA context and GPU allocations. Testing it would mean moving it out of that function, a source change outside this test and docs PR.
Very large NCCL_GIN_PROXY_POLL_BATCH wraps negative: real, but it is in existing source ((int)pollBatchParam with no upper clamp). The fix belongs in gin_host_proxy.cc, so I'll leave it for a separate source change.
ops == 256: the process runs only this gtest, and every run reported exactly 256 data ops on rank 0. The exact match is deliberate: other ops reaching the plugin should fail the test, and the hinted-op bound is computed from 256.
Log-line regex: a reworded line fails loudly at the RMA example did not report assert rather than silently dropping coverage, and the test sets NCCL_DEBUG_SUBSYS=INIT,NET itself.
Shared run-and-log helper: about 12 lines in common with the existing _run_all_reduce; left for a later cleanup to keep this PR's scope small.
ext_rma off in CI: agreed. Enabling the marker is separate, in-progress work.

amd-argus

ARGUS

Findings

1 finding(s) that could not be posted as inline comments

projects/rccl/docs/how-to/using-rccl-rma-plugin-api.rst:81 (Low) The page's description of plugins/rma/example was not updated for the two behaviours this PR gives it At using-rccl-rma-plugin-api.rst:81-83, outside this PR's hunks, the example is still described as only demonstrating "the lifecycle, registration and data-operation callback signatures". It now also reports NCCL_NET_DEVICE_GIN_PROXY so NCCL_GIN_TYPE=2 selects it (plugin.c:89), and counts hinted data ops and logs them at finalize (plugin.c:167-176), which is the line the new _RMA_EXAMPLE_COUNTS regex depends on (test_rma_plugin.py:293). Since this PR already edits this page, worth a sentence there?

Bottom line: Worth settling before approval: with NCCL_GIN_PLUGIN also set, the RMA example's new NCCL_NET_DEVICE_GIN_PROXY type makes gin.cc skip the built-in proxy, so test_rma_example_backs_the_gin_proxy no longer proves its claim.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

rahulvaidya20 · 2026-10-11T19:00:51Z

Third Argus round:

GIN proxy registration (f9a5541943): the built-in proxy is not skipped. After the external GIN example is skipped it is finalized and, with the default NCCL_GIN_PLUGIN_REF_COUNT=0, unloaded, which resets its state to LoadReady (gin.cc:81). The dedupe loop only compares against plugins at InitReady or above, so the built-in proxy is registered. The test now also asserts that neither GIN type previously assigned nor Failed to initialize any GIN plugin appears in the log.
Block guard (f236f9dc66): removed. The kernel comment now states the single-CTA launch once.
Suite name (a1f7b63bd0): now names both gin_host.cc and gin_host_proxy.cc.
Example description (b7315b4820): the RMA plugin how-to now says the example reports the GIN proxy device type and logs data and hinted op counts at finalize.

Not changed:

Counts across communicators: the counters are documented as process-wide, and the test runs one communicator. Resetting at finalize would still mix concurrent communicators; true per-communicator counts need plumbing from each request back to its communicator, which is more than the example needs.
Counter path in the host fixture: counter updates are outside batching and the hint. PutBurst_LocalCounter covers them end to end by waiting for all 256 counter increments.

Re-validated on MI350X (gfx950, RoCE, ROCm 10.1): host microtests 26/26, ext-rma 6/6 (hinted ops 0, 192 and 248 for batches 1, 4 and 32), 20 of 20 default-batch runs at 248, and PutBurst_LocalCounter plus WaitCounter_Local over the built-in IB verbs GIN proxy.

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

dlamd1dai

The host microtests provide good coverage of batching, hint propagation, flush termination, peer isolation, and slot release. I found no additional correctness or contract defects.

Requesting changes because the ext-RMA test still asserts a timing-dependent positive hint count. Once the existing thread is resolved, this looks ready from a code-review perspective. CI is not currently clean: host tests pass, while MCI extra and TheRock are failing and RCCL precheckin is pending.

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

dlamd1dai

The ext-RMA test no longer requires a positive hinted-operation count. A batch of one still requires zero hinted operations because its upper bound is zero, and larger batches now assert only the safe upper bound. The host microtests continue to lock the exact hint pattern.

No remaining findings. Approving from a code-review perspective.

JIRA ID: AICOMRCCL-2807 ## What Host-only microtests for `src/gin/gin_host_proxy.cc` and the versioned GPU-context writers in `src/gin/proxy_gpucontext/` (`gpucontext{,_v1,_v2}.cc`), in a new `rccl-UnitTestsMicroGinHostProxy` binary (65 tests where rocmwrap.h's DMA-BUF export gate is on, 61 where it is off). Overlaps rahulvaidya20's #13221 (same file name, same unit); will reconcile with it. ## How - `gin-host-proxy-test.cc` `#include`s the four hipified units and seams the RMA plugin and the device allocators in the TU; `allocMemCPUAccessible` is the real `gdrwrap.h` helper, steered by `ncclGdrCopy`. `hipHostMalloc`, `hipMemset` and the DMA-BUF export go through `fakes/hip_fakes.cc` hooks (the export hook is new, built where CMake finds `hipMemGetHandleForAddressRange`), and logging and params come from `fakes/nccl_fakes.cc`. GFDs are posted straight into the host-side queues. - Pins 3 production bugs as observed, not fixed (AICOMRCCL-2991..2993): - `ncclGinProxyRegMrSym`: when both DMA-BUF registrations fail, it returns before closing the second exported fd. - `ncclGinProxyCreateContext`: accepts `backendVersion` 0, fails in the GPU-context writer, and releases nothing it created. - `ncclGinProxyDestroyContext` never frees the signal-offsets buffer. ## Coverage Lines: `gin_host_proxy.cc` 476/476 with the export gate on (456/467 with it off: the 11 missed lines are the DMA-BUF registration, unreachable there), `gpucontext.cc` 33/33, `gpucontext_v1.cc` 10/10, `gpucontext_v2.cc` 13/13. ## Mutation Testing 57/57 killed: 53 in `gin_host_proxy.cc`, 4 in the GPU-context writers. Review follow-up: 15 more (allocator pairing, `forceHost` in `gdrwrap.h`, plugin properties, queue-size branches, the per-qword wait), all killed; dropping the plugin's DMA-BUF flag is caught only with the gate on, since with it off the flag changes nothing. ## Test Plan Standalone build (HIP 7.1, gate off): 61/61; with the gate forced on locally: 65/65. `--gtest_shuffle --gtest_repeat=50` is clean in both. --------- Co-authored-by: Claude <noreply@anthropic.com>

…mple

…oxy host microtests

peymanr · 2026-11-05T00:45:52Z

PR Pundit review

(Testing PR Pundit Code Review agent. Not all comments may be relevant. This is a one-time automated pass: please address what is useful, no need to request a re-review. Please reply to each point, whether it applies or not: we harvest these replies, and both kinds are enormously helpful for tuning the agent. Reach out to @peymanr for suggestions.)

#		Finding	Where
1	🟠	PutBurst_LocalCounter passes despite corrupted or misdirected remote writes: this PR verifies only local completion	`projects/rccl/test/transport/GinDeviceMPITests.cpp:1886-1965`

1. 🟠 Should fix: PutBurst_LocalCounter passes despite corrupted or misdirected remote writes: this PR verifies only local completion

projects/rccl/test/transport/GinDeviceMPITests.cpp:1886-1965

PutBurst_LocalCounter only checks that rank 0's local counter reaches nPuts within 60 s. dSrc is never initialized, all 256 puts write dst offset 0, and rank 1 reads nothing back. On the IB/RoCE configs that add it, the test therefore proves only that the burst completes without hanging. A remote-side check of distinct offsets or the peer counter would catch dropped or corrupted writes. This is by design for the no-data plugins/rma/example run, and the counter provisioning (ginCounterCount = 2 for index 1, railGinBarrierCount = 1) follows this file's convention.

Who is affected: CI runs of GinMPIDeviceTests.PutBurst_LocalCounter on the mi300 IB and mi355x RoCE configs, with and without NCCL_GIN_PROXY_POLL_BATCH=1.

Checked against the code by an independent pass.

Related past reviews: #10388 on projects/rccl/test/transport/GinDeviceMPITests.cpp, #11089 on projects/rccl/test/transport/GinDeviceMPITests.cpp, #9188 on projects/rccl/test/transport/GinDeviceMPITests.cpp, #12640 on projects/rccl/test/device/GinDeviceTests.cpp

_{Shared from a pilot of PR Pundit, an agentic code reviewer that builds and tests the change and reads the project's past reviews. Reviewed at cc4da33a; findings checked and posted by @peymanr; points already raised on this PR are left out.}

amd-argus

ARGUS

Bottom line: LGTM.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

github-actions · 2026-11-05T01:10:20Z

pr-rocm-systems-13221-oci-rccl-pr-smoke

View this job on CoCo

Builds

Stack	Component	Version	Commit	Status
pr-13221_rccl-perf_rocm-10.1-ga	gcc-13.2	13.2.0	13.2.0	cached
pr-13221_rccl-perf_rocm-10.1-ga	rocm-10.1-ga	10.1.0	10.1.0	cached
pr-13221_rccl-perf_rocm-10.1-ga	openmpi-5.0.8	5.0.8	5.0.8	cached
pr-13221_rccl-perf_rocm-10.1-ga	rccl-pr-13221	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-perf_rocm-10.1-ga	rccl-tests-develop-perf	develop	`8de4115`	cached
pr-13221_rccl-perf_rocm-10.2-latest	gcc-13.2	13.2.0	13.2.0	cached
pr-13221_rccl-perf_rocm-10.2-latest	rocm-10.2-latest	10.2.0rc1	10.2.0rc1	cached
pr-13221_rccl-perf_rocm-10.2-latest	openmpi-5.0.8	5.0.8	5.0.8	cached
pr-13221_rccl-perf_rocm-10.2-latest	rccl-pr-13221	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-perf_rocm-10.2-latest	rccl-tests-develop-perf	develop	`8de4115`	cached
pr-13221_rccl-smoke_rocm-10.1-ga	gcc-13.2	13.2.0	13.2.0	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rocm-10.1-ga	10.1.0	10.1.0	cached
pr-13221_rccl-smoke_rocm-10.1-ga	openmpi-5.0.8	5.0.8	5.0.8	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rccl-pr-13221-smoke	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rccl_gtest-pr-13221	`cc4da33`	`cc4da33`	cached
rocm-10.1-ga_rccl-develop-perf	gcc-13.2	13.2.0	13.2.0	cached
rocm-10.1-ga_rccl-develop-perf	rocm-10.1-ga	10.1.0	10.1.0	cached
rocm-10.1-ga_rccl-develop-perf	openmpi-5.0.8	5.0.8	5.0.8	cached
rocm-10.1-ga_rccl-develop-perf	rccl-develop-f030609e6984	`f030609`	`f030609`	cached
rocm-10.1-ga_rccl-develop-perf	rccl-tests-develop-perf	develop	`8de4115`	cached

Failures

No failing tests in this run.

15 tests skipped.

Performance regressions

`pr-13221_rccl-perf_rocm-10.1-ga` vs `rocm-10.1-ga_rccl-develop-perf`

Readings: flat 2070.

No confirmed regressions or improvements.

github-actions · 2026-11-05T01:10:21Z

pr-rocm-systems-13221-ruby64-rccl-pr-smoke

View this job on CoCo

Builds

Stack	Component	Version	Commit	Status
pr-13221_rccl-perf_rocm-10.1-ga	gcc-11.5	11.5.0	11.5.0	cached
pr-13221_rccl-perf_rocm-10.1-ga	rocm-10.1-ga	10.1.0	10.1.0	cached
pr-13221_rccl-perf_rocm-10.1-ga	openmpi-4.0.2	4.0.2rc3	4.0.2rc3	cached
pr-13221_rccl-perf_rocm-10.1-ga	rccl-pr-13221	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-perf_rocm-10.1-ga	rcclt-develop-perf	develop	`8de4115`	cached
pr-13221_rccl-perf_rocm-10.2-latest	gcc-11.5	11.5.0	11.5.0	cached
pr-13221_rccl-perf_rocm-10.2-latest	rocm-10.2-latest	10.2.0rc1	10.2.0rc1	cached
pr-13221_rccl-perf_rocm-10.2-latest	openmpi-4.0.2	4.0.2rc3	4.0.2rc3	cached
pr-13221_rccl-perf_rocm-10.2-latest	rccl-pr-13221	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-perf_rocm-10.2-latest	rcclt-develop-perf	develop	`8de4115`	cached
pr-13221_rccl-smoke_rocm-10.1-ga	gcc-11.5	11.5.0	11.5.0	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rocm-10.1-ga	10.1.0	10.1.0	cached
pr-13221_rccl-smoke_rocm-10.1-ga	openmpi-4.0.2	4.0.2rc3	4.0.2rc3	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rccl-pr-13221-smoke	`cc4da33`	`cc4da33`	cached
pr-13221_rccl-smoke_rocm-10.1-ga	rccl_gtest-pr-13221	`cc4da33`	`cc4da33`	cached
rocm-10.1-ga_rccl-develop-perf	gcc-11.5	11.5.0	11.5.0	cached
rocm-10.1-ga_rccl-develop-perf	rocm-10.1-ga	10.1.0	10.1.0	cached
rocm-10.1-ga_rccl-develop-perf	openmpi-4.0.2	4.0.2rc3	4.0.2rc3	cached
rocm-10.1-ga_rccl-develop-perf	rccl-develop-f030609e6984	`f030609`	`f030609`	cached
rocm-10.1-ga_rccl-develop-perf	rcclt-develop-perf	develop	`8de4115`	cached

Failures

No failing tests in this run.

15 tests skipped.

Performance regressions

`pr-13221_rccl-perf_rocm-10.1-ga` vs `rocm-10.1-ga_rccl-develop-perf`

Readings: flat 2070.

Fails the run: 0 regressed point(s) past 8%, each confirmed by an adjacent size.

No confirmed regressions or improvements.

systems-assistant · 2026-11-08T06:00:36Z

TheRock Submodule Bump Activity

Newest first

2026-11-08 — Included in TheRock via ROCm/TheRock#9793.

rahulvaidya20 requested review from a team and a balanced review from Copilot October 9, 2026 22:46

github-actions Bot added documentation Improvements or additions to documentation project: rccl labels Oct 9, 2026

Copilot started reviewing on behalf of rahulvaidya20 October 9, 2026 22:47

View session

Copilot AI reviewed Oct 9, 2026

View reviewed changes

systems-assistant Bot added the organization: ROCm label Oct 9, 2026

amd-argus Bot reviewed Oct 9, 2026

View reviewed changes

amd-argus Bot reviewed Oct 10, 2026

View reviewed changes

amd-argus Bot reviewed Oct 11, 2026

View reviewed changes

dlamd1dai reviewed Oct 12, 2026

View reviewed changes

Comment thread projects/rccl/plugins/rma/example/plugin.c Resolved

Comment thread projects/rccl/plugins/rma/example/nccl/common.h Resolved

amd-argus Bot reviewed Oct 12, 2026

View reviewed changes

dlamd1dai reviewed Oct 13, 2026

View reviewed changes

Comment thread projects/rccl/test/transport/GinDeviceMPITests.cpp Resolved

Comment thread projects/rccl/test/ext-plugins/tests/ext-rma/test_rma_plugin.py Resolved

Comment thread projects/rccl/docs/how-to/using-rccl-rma-plugin-api.rst Resolved

amd-argus Bot reviewed Oct 14, 2026

View reviewed changes

dlamd1dai reviewed Oct 15, 2026

View reviewed changes

Comment thread projects/rccl/tools/scripts/test_runner/configs/mi300_mellanox_ib_func.json Resolved

aduforest mentioned this pull request Oct 15, 2026

test(rccl): add host microtests for the GIN host proxy #13596

Merged

amd-argus Bot reviewed Oct 16, 2026

View reviewed changes

dlamd1dai requested changes Oct 16, 2026

View reviewed changes

amd-argus Bot reviewed Oct 16, 2026

View reviewed changes

dlamd1dai approved these changes Oct 16, 2026

View reviewed changes

pvallem approved these changes Oct 16, 2026

View reviewed changes

rahulvaidya20 force-pushed the users/rahulvaidya20/aicomrccl-2018-gin-proxy-batching branch 3 times, most recently from 45b0bb6 to 989789c Compare October 21, 2026 02:25

abouteiller approved these changes Oct 27, 2026

View reviewed changes

rahulvaidya20 added 6 commits November 4, 2026 16:00

Report GIN proxy device type and count aggregated ops in the RMA example

ba532e5

Add GIN put burst test that waits only on local completion

2b334d4

Check GIN proxy batching and the aggregation hint through the RMA exa…

cba857d

…mple

Cover the aggregation hint on signal, get and flush ops in the GIN pr…

357bf33

…oxy host microtests

Run the GIN put burst in the MI300X and MI355X functional configs

6cf2dd1

Document NCCL_GIN_PROXY_POLL_BATCH and the RMA aggregation hint

cc4da33

rahulvaidya20 force-pushed the users/rahulvaidya20/aicomrccl-2018-gin-proxy-batching branch from 989789c to cc4da33 Compare November 5, 2026 00:24

amd-argus Bot reviewed Nov 5, 2026

View reviewed changes

rahulvaidya20 merged commit 1b84bd2 into develop Nov 5, 2026
67 of 71 checks passed

rahulvaidya20 deleted the users/rahulvaidya20/aicomrccl-2018-gin-proxy-batching branch November 5, 2026 16:24

Conversation

rahulvaidya20 commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Issue Tracking

Test Plan

Test Result

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Oct 9, 2026

✅ All Policy Checks Passed

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟢 Approval recommended

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Findings

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

rahulvaidya20 commented Oct 10, 2026

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

rahulvaidya20 commented Oct 11, 2026

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Findings

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

rahulvaidya20 commented Oct 11, 2026

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dlamd1dai left a comment

Choose a reason for hiding this comment

Uh oh!

amd-argus Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dlamd1dai left a comment

rahulvaidya20 commented Oct 9, 2026 •

edited

Loading

`pr-13221_rccl-perf_rocm-10.1-ga` vs `rocm-10.1-ga_rccl-develop-perf`

`pr-13221_rccl-perf_rocm-10.1-ga` vs `rocm-10.1-ga_rccl-develop-perf`